Optimal Control of Voltage-Forming Grid Inverters by Model Predictive Control and Reinforcement Learning

Optimal control approaches, such as model predictive control (MPC) and reinforcement learning (RL), are promising for decentralized grid applications but face challenges related to model inaccuracies and safety constraints. This work experimentally compares both methods on industrial-grade three-phase, three-level inverter hardware with an LC-filter and common neutral point, going beyond simulation-based studies. Under identical safety constraints, both controllers show similarly fast transient settling. In steady state, MPC achieves a normalized mean absolute error (with respect to the reference voltage amplitude) of 0.33%, while the RL controller reaches 0.18%. These results demonstrate that both model-based and model-free control can attain optimal performance when subjected to real-world and real-time conditions. Finally, this contribution highlights the key challenges in transferring theoretical concepts to practical power electronic inverters and offers a detailed implementation guide for both researchers and industrial practioners.